Towards Parameter-Free Classification of Sound Effects in Movies

Towards Parameter-Free Classification of Sound Effects in Movies Selina Chu, Shrikanth Narayanan *, C.-C Jay Kuo * Department of Computer Science * Department of Electrical Engineering University of Southern California, Los Angeles, CA, USA 90089 ABSTRACT The problem of identifying intense events via multimedia data mining in films is investigated in this work. Movies are mainly characterized by dialog, music, and sound effects. We begin our investigation with detecting interesting events through sound effects. Sound effects are neither speech nor music, but are closely associated with interesting events such as car chases and gun shots. In this work, we utilize low-level audio features including MFCC and energy to identify sound effects. It was shown in previous work that the Hidden Markov model (HMM) works well for speech/audio signals. However, this technique requires a careful choice in designing the model and choosing correct parameters. In this work, we introduce a framework that will avoid such necessity and works well with semi- and nonparametric learning algorithms. Keywords: Multimedia mining, data mining, audio classification, pattern recognition, video mining. 1. INTRODUCTION Content-based analysis and video mining provides us with ways for indexing multimedia databases that will facilitate easier and more meaningful search and query. Past research in multimedia content analysis has focused mostly on specific events, like sports [1, 2, 3, 4] or news broadcasts [5]. Most focus on using indicators such as the excited voice of the commentator and cheering of the audience. Little work has been done on actual classification and mining of movies. Much research interest in the movie analysis has been on event detection, boundary detection or event segmentation, but not necessarily identifying an interesting event. The problem of identifying intense events via data mining in films is investigated in this work. Movies are mainly characterized by dialog, music, and sound effects. Most previous work focuses on obtaining summarization or compact representations of movies using the speech and music information. For example, audio features were used in [6] to detect events in action films, without distinguishing among explosions and shots. The same group had previously looked at genre classification, but using very different classes such as news broadcast, sports-car races, sports-tennis, commercials, and cartoons [5]. In [7], audio-visual cues were used, such as sound changes from the market scene to the conversation, for audio transitions, and lighting and colors for video. All these were obtained manually to obtain their audio and video transitions in deriving models for analyzing movies. We begin our investigation on detecting interesting events through sound effects. While speech can be abstracted from the subtitles, sound effects are typically detected by listening. Due to the disparate nature of sound effects, it is not too clear how to describe them. Sound effects are considered neither speech nor music, but a combination of some specific audio signals. One possible way to model the sound effects is the use of hidden Markov models (HMM). However, as explained in Sec. 2, we observe some problems in applying HMM to audio sound effects. Furthermore, it is desirable to find a way to utilize simple learning algorithms, without resorting to complicated training procedures. These are the limitations and difficulties that motivate our work. The intuition behind our framework is to devise a simpler way of identifying audio sound effects without the hassle of fine-tuning and uses as few parameters as possible or even none at all. The advantage of our framework is that we are not restricted to using one classifier; one can plug in any appropriate classifier that is deemed suitable for the particular class of audio.

The rest of the paper is organized as follows. Sec. 2 reviews previous related work. Sec. 3 explains the observations we made in using HMMs. In Sec. 4, we describe our framework for classifying sound effects. In Sec. 5 we experimentally compare various classifiers using our framework. Finally, Sec. 6 contains our conclusion and directions for future work. 2. REVIEW OF PREVIOUS WORK Rajapakse and Wyse [13] proposed a hybrid model comprised of GMMs and HMMs. It is used to model generic sounds with large intra class perceptual variations. All training sequences are sent through this hybrid model to estimate the model parameters that minimize the combined classifier error rate. This is motivated by the observation that GMM can capture the structural information while HMM is used to extract the transition information in sequences. Zhang and Kuo [7] introduced a method for automatic segmentation and classification of audiovisual data based on audio features. Classification is performed for discriminating between basic types of sounds, including speech with or without music background, music, song, environmental sound with or without music background, silence, etc. This model does not rely on the learning algorithm, but rather uses rule-based heuristics to segment and classify the audio signals. They utilize a back-to-back window scheme to analyze and compare audio features such as energy, zerocrossing and fundamental frequency for segmentation and classification of the segmented events. Pfeiffer et al. [1] described a framework for audio content analysis, where they also performed classification and segmentation. It is a two-stage process. The first stage is audio classification, where audio signals are separated into music, speech, silence and other sounds. The second stage is used for segmentation in determining the syllable, word or sentence boundaries for speech, the note bar or theme boundaries. They use features such as the amplitude, frequency, pitch, onset, offset, and frequency transitions to perform comparison for changes and apply thresholding for classification and segmentation. Detection of sounds such as gunshots, cries and explosions was used to indicate violence in their audio data. 3. SOUND EFFECT CLASSIFICATION WITH HMM A well known solution to audio-related problems is based on Hidden Markov models (HMM) and Gaussian Mixture models (GMM), both of which have been well studied in the audio signal processing field [9, 10, 11, 12, 13]. Because of the success HMM has attained in speech recognition, it would seem to be the classifier of choice. Observing results given in Table 1, it is indeed demonstrating HMM to perform better than other classifiers under comparison. Another advantage HMM has over other classifiers in comparison is its ability to accommodate the time variation characteristics. The main drawback of HMM is the non-trivial nature in the design, and it is even more stringent when making changes. Designing an HMM requires expertise in this area. It is with experience in using HMM for a particular domain that one would gain insight into how many states it should contain and/or where to place transitions. HMM is a robust classifier once it is designed, but vigorous and complicated when tuning to achieve an optimal model. Other consideration that one needs to take into account is the amount of data. Because increasing the number of transitions and states translate to increasing the degrees of freedom, in order to achieve a stable system, we would also need an increasing amount of data. Unlike speech, there is no plethora of sound effect data from movies, since only a very limited number of instances of each kind of sound really exist in a movie. Even with the complexity of HMM, we are achieving a classification rate of 87%. Technique Classification accuracy % Hidden Markov Model (HMM) 87.138 Gaussian Mixture Model (GMM) 70.899 K Nearest Neighbor (KNN) 83.340 Naïve Bayes Classifier 74.516 Table 1: Comparison of classifiers on a 4-class classification task of sound effects (similar to experiments in 5.4), but was performed without dimensionality reduction or the preprocessing stages introduced in this work.

We begin our research on evaluating the use of HMM for modeling sound effects. The HMM created for this experiment is a left-to-right model, and the number of states for the HMM is set to 3. We have tried several different numbers of states, but found three states to be sufficient empirically. The distribution of observations was modeled by four Gaussian mixtures. Again, we have analyzed different number of Gaussian mixtures and found 4-5 mixtures to produce good results. For brevity, we will not explain the details of HMM, but direct the interested reader to [4] for a complete description. In addition to the 3-state HMM for each sound class, there is also an HMM for the silence model, which is similar to the HMM in Fig. 1 for each class except that it has extra transitions from states 1 to 3 and from 3 to 1. It is used to make the model more robust by allowing states to absorb noise in the data. Figure 1: Illustration of the 3-stage forward HMM. We compared classification in discriminating between four sound effects; namely, explosion, gunshot, glass-shattering and screaming. The features, data and classes are detailed in Sec. 4. We examined two types of HMM: the 3-state forward-type and the 3-state ergodic-type. Ergodic HMM means that every state can be reached from any other state. In the forward-type, each state only has transition probabilities going to itself to the next successor state. We used MFCC as features, along with a silence model to absorb noise in the training data. The overall accuracy rate for this four-class classification task is given in Table 2. Types of HMM Classification Accuracy % 3-State Forward 87.1 3-State Ergodic 90.7 Table 2: Classification of HMM, comparing between 3-stage forward and 3-stage ergodic HMMs. Another possible problem with using HMM could be the fact that the duration of each audio clip differs widely. The duration of actual sounds also varies widely between clips. For example, there could be a really long scream and the rest could be much shorter. These might be possible to fix by different construction of the models. As mentioned earlier, it is necessary to fine-tune the system to achieve a near optimal model. It is with this tedious and complicated design process that has let us to examine other means of classifying sound effects. 4. PROPOSED SEMI-PARAMETRIC MODEL FOR AUDIO CLASSIFICATION In this section, we introduce a new framework for classifying sound effects using semi-parametric or nonparametric classifiers depending on the choice of users. The advantage of this framework is that we are not restricted to any particular classifier; one can plug in any classifier the designer chooses. 4.1 Proposed Framework We begin by segmenting the sound effect clips into n sections based on some segmentation criteria. In this work, we will examine two different criteria. The first one is to simply to divide audio clips recursively with equal-lengths. Since there is no obvious segmentation as to how a sound should be segmented, we begin by separating it into two segments, and then for each segment, we sub-divide them further. We repeat the process for three levels, resulting in at most eight segments. The second one is to segment the data based on the amount of energy each section contains. Instead of

having equal-length segments, we use equal-energy segments. This means that each section should contain the same, or very similar, amount of energy. Since we are working with audio, we can only analyze the data based on the sampling rate. Unless we set the frame to be very small, it is difficult to divide the data into segments containing an equal amount of energy. But each segment should contain, more or less, similar amount of energy. The intuition behind both methods is to provide a robust way of alleviating the need to predetermine the length of each data sequence and to adjust for the variation in each data sample. Since the speed of sound patterns vary, dividing the audio clip solely based on the time might not be enough to accommodate for such variation. Once segments are found, we can extract the corresponding MFCCs from each segment to be used as features. Segments belonging to that section are placed together. To make this clearer, we will look at an example. Let us say that a sound clip is divided into two segments. The first segment requires five sampling frames, using the specification of 20 msec per frame, while the second segment requires ten. Each frame will then contain 40 MFCCs. Each featureframe of the first segment will be placed together with other feature-frames from the first segment of other sound clips. Each frame of features can be viewed as an individual data sequence used to train the classifier for the first segment. In other words, we combine every frame of the first segment from all the training sound clips together, noting only the corresponding labels, but ignoring the origin of each specific sound clip. We then use these features as input to the classifier. This way, each section will have its own classifier. For example, if we divide all the sound clips into 2 sections, then there should be two separate classifiers in total. We repeat the feature extraction process for the testing data. Before any classification is performed, the features for both training and testing go through a dimension reduction process. Then, the user can employ any classifier that he or she chooses. As we will see from our experiments, even using simple classifiers, such as KNN or Naïve Bayes, is sufficient in producing reasonable results. Procedure for Training: Step 1: Divide the data sequence into n segments, based either on time (1st case) and energy (2nd case) Step 2: For each of the n segments, obtain the MFCC features for each sample frame Step 3: Combine all feature-frames of one section Step 4: Reduce the dimensionality of the dataset Step 5: Train classifier of choice with the dimension-reduced data Procedure for Testing: Step 1: the same as Step 1 given above Step 2: the same as Step 2 given above Step 3: Reduce dimensionality of data Step 4: For each feature-frame of section n, classify that frame using the chosen classifier corresponding to section n Step 5: When all feature-frames of one querying sound effect clip are classified, each frame will perform majority voting to determine the class of each section. Step 6: Each section will also vote and majority wins. (If there are more than 2 sections, we ignore the beginning and end sections.) Ties are broken by giving the vote to the section with more frames. If they have the same number of frames, we randomly pick between the two. Repeat Step 4 until all queries are processed Table 3: Summary of our framework for classification To perform a particular query on a new instance of sound effect, we divide the data into segments based on the segmentation criteria and extract the MFCC features similar to how the training data was processed. However, in this case, we keep track of the sound clip that each feature-frame originates from. Then, we treat each feature-frame as an individual data instance to be query. Each individual data sequence will be classified accordingly, and each section will then use a majority voting to decide its classification. Once each section has obtained its classification, we allow the significant sections to vote. What we mean by significant is that if there are more than 2 sections, we will ignore

the first and last sections and only use the middle ones. The reason behind this is that the beginning and end usually consists of similar patterns, which is from silent or low volume to a sudden surge in noise and vice versa for the ending sections. If there are ties, we pick the one with more frames as the winning section. If they are the same, we randomly pick between the two choices. 5. EXPERIMENTAL PROCEDURE AND EVALUATION In this section, we conduct a number of experiments to compare the classification using our framework with various classifiers. 5.1 Data Set and Feature Extraction The audio data samples collected are mono-channel, 16 bits per sample with a sampling rate of 16 khz and are of varying length. We use data from four classes of sound effects. They are: explosion, glass shattering, gunshot and screaming. There are 100 samples from each of the four classes. The duration of each audio segment ranges from 3 to 10 seconds. Features are analyzed and extracted for every 20ms frame. We used energy and Mel-Frequency Cepstral Coefficients (MFCC) to the 40th order as features in our classification tasks. MFCC has been shown to work well as features for classification [3, 13] and to discriminate between classes such as silence, applause, ball-hitting, cheering, and music [3]. 80% of the data is used for training and 20% for testing. These settings are applied to every experiment performed in this paper. All data, including querying sequences, are normalized to zero mean and unit variance. 5.2 Dimensionality Reduction Due to the inherently high dimensionality nature of the data, it is important that we reduce the dimensions as a preprocessing step. For example, each audio clip is about 3-10 seconds in duration. Since we use an analysis window of 20 msec frame to extract energy and MFCC features, where each MFCC frame is made up of 40 coefficients, and each audio clip will generate between 150 to 500 frames. As shown in Table 4, the decrease in the amount of time required for 40 dimensions to 12 dimensions is about two folds. Therefore reduction in dimensionality of the data is important in practice for complexity reduction. As demonstrated in Fig. 4, reducing the dimensions to 5 or 6 will produce results similar to using the features before dimensionality reduction. 40 dimensions 12 dimensions K-Nearest Neighbors 16.42 min 6.42 min GMM 10.52 min 4.8 min Naïve Bayes 0.1 min 0.1 min Table 4: Comparison of the CPU time (minutes) used for classification A simple and useful dimensionality reduction technique for multidimensional data is principal component analysis (PCA). PCA reduces the dimensionality of data by eliminating redundancy between the dimensions. It works by reducing the dimensionality based on correlation. It tries to collapse correlated dimensions, while leaving the uncorrelated ones intact. The correlations between the reduced dimensions should be close to zero since this will provide us with independent information on the data. Performing PCA is similar to performing singular value decomposition on the covariance matrix of the data. Detail on PCA can be found in [16] for a more detailed treatment. 5.3 Classification Algorithms In this section, we will briefly review the classifiers used in our experiments. 5.3.1 Gaussian Mixture Models

With GMMs [17], each data class is modeled as a mixture of several Gaussian clusters. Mixture models are semiparametric. Each mixture component is a Gaussian and is represented by the mean and covariance matrix of the data. Once the model is generated, conditional probabilities can be computed using k m = k p( x X ) p( x j) P( j), j= 1 where m k denotes the number of components, P(j) is the prior probability that data x was generated by component j, and P(x j) is the mixture component density. EM algorithm [14] is then used to find the maximum likelihood parameters of each class. We want to find a model C k that optimizes the posterior probability given p(c k x). The maximum likelihood criteria, given equally likely prior probabilities P(Ck) for all classes, is C = arg max p(x C k ). C = arg max p( x C ). GMM k 5.3.2 K-Nearest Neighbors The K-nearest neighbor (KNN) is a simple supervised learning algorithm where a new query is classified based on the majority class of its k nearest neighbors. The nearest neighbors are defined by the minimum distance from the query to the training samples. A commonly used distance measure is the Euclidean distance, n 2 d( x, y) = ( x i y i ). i= 1 This is the distance measure used with all KNN related experiments in this paper. Please refer to [15] for a more thorough explanation of this technique. 5.3.3 Naïve Bayes Naïve Bayes (NB) classifier is a supervised learning algorithm that makes the assumption of conditional independence of the attributes given the class. In other words, each feature or data point is probabilistically independent of all other data points. The probability of observing conjunctions x 1, x 2 x n is the product of the probabilities for individual data points: P ( x1, x2 L xn ck ) = P( xi ck ) i. The naïve Bayes classifier is simply c = arg max P( c ) P( x c ) NB c C j i, where p(c j ) is the prior, which is simply counting the frequency with which each x j occurs in the training data. A more detailed description of this technique can be found in [15] 5.4 Experimental Evaluation j k i j To demonstrate the effectiveness of our system, we compare classification results with three different types of classifiers, including Gaussian mixture models, K-nearest neighbors, and Naïve Bayes Classifier as described in the previous section. Again, we are discriminating between four classes of sound effects: explosion, glass shattering, gunshot, and screaming. For KNN, we utilize the 10-nearest neighbor queries to obtain the results, and for GMM, we set the number of mixtures to 5. Since GMM models each class, we have 4 different GMM models for each section. For example, if we decide to use two segments for segmentation, then there are 8 models all together. Finally, no specific settings are required for Naïve Bayes.

The hardware used for all experiments was a Pentium IV 2GHz with 512 MB RAM. We used Matlab for all implementations. Since Matlab is an interpreted language, the overall time is slow. However, for time-wise comparison, we are only interested in the relative performance. We observed experiments for different settings: 1) Variation in dividing into N equal-length segments 2) Equal-length versus Equal-energy as segmentation criteria 3) Effects of PCA on K dimension Figure 2: Comparison of classification accuracy between GMM and KNN, using n equal-length segments (n = 1, 2, 4, and 8), along with PCA to reduce from 40 dimensions. In the first experiment, results are acquired by varying the number of segments used to divide the audio data. We investigate the equal-length segmentation for the following conditions: no segmentation, 2 segments, 4 segments, and 8 segments. The results are shown in Fig. 2. In both instances, the 4-segment case has better performance than the 8- segment case. For the second experiment, we explore the disparity between equal-length and equal-energy segmentation. Since KNN and NB produce similar results, we make comparison between KNN and GMM in Fig. 3. For both classifiers, using equal-energy as the segmentation criteria produces the best result, with KNN achieving about 91% for 5-12 dimensions. The previous experiment indicates that equal-energy segmentation yields better results in all cases. Thus, for the last experiment, we only examine results for equal-energy segmentation. We compared the accuracy between the three classifiers, while varying the dimensions being reduced. The results are shown in Fig. 4. An interesting finding is that it is possible to obtain adequate results even when the dimensions are reduced to as little as 5 or 6. We also observed that for higher dimensions, Naïve Bayes can achieve a classification rate of 93% when using 34 dimensions. (One should note that Naïve Bayes is very simple and can be computed very quickly. Using 34 dimensions for NB is still faster than KNN or GMM applied to one dimensional data.) Overall, KNN performed the best; it can provide us with a stable classification. The accuracy only drops below 90% for 4 or lower dimensions. Finally, GMM has proven to yield the worst performance in all three experiments.

Figure 3: Comparison of classification rate between using equal-length and equal-energy for segmentation, where data are reduced using PCA from 40 MFCC to 1-12 dimensions. Figure 4: Comparison of classification rates for different dimensions using three different classifiers -- 10-nearest neighbors (10-NN), Naïve Bayes (NB), and Gaussian mixture models (GMM) with 5 mixtures. Data was reduced from 40 MFCC to 1-39 dimensions. 6. CONCLUSION AND FUTURE WORK In this work, we introduced a framework that exploits the idea of obviating the need for parametric algorithms, where users are required to fine-tune their parameters to achieve an optimal classification result. We utilize the idea of using semi- or non-parametric classifiers, along with the dimensionality reduction technique, to produce an improvement in classification accuracy of sound effects. We have also demonstrated a method to segment each audio data sequence by using energy as an indicator. We have shown our system to perform better than HMMs without a need of fine-tuning any parameters nor requiring an expert to design and use this system. The reason why HMM works well with speech is that words are easily divided up by their phonemes, and each phoneme can be modeled by a single state of an HMM.

Sound effect, on the other hand, has no clear boundaries; the duration varies widely between different audio clips of the same type of sound. Therefore, we can not treat sound effects similar to speech. Our system is still in its infancy stage. Future work includes automating the hierarchical segmentation of the data. Currently, we are sub-dividing the data up to three levels. It is possible to use very low dimensions of the data, to quickly determine the number of segments we should utilize, and then refine our model using higher dimensions. Another possible direction for future work is to make sections more adaptive using the notion of relevance feedback and weights. If a section continually misclassifies, we should reduce the weight, or significance, of this section and increase the weight for sections that produce better performance. REFERENCES 1. S. Pfeiffer, S. Fischer and W. Effelsberg, Automatic audio content analysis, Praktische Informatik IV, Univ. Mannheim, Mannheim, Germany, 1996. 2. R. Lienhart, S. Pfeiffer, S. Fischer, Automatic movie abstracting, Technical Report TR-97-003, Praktische Informatik IV, University of Mannheim, July, 1997 3. R. Radhakrishan, Z. Xiong, A. Divakaran, and Y. Ishikawa, Generation of sports highlights using a combination of supervised & unsupervised learning in audio domain, International Conference on Pacific Rim Conference on Multimedia, Vol. 2, pp. 935-939, December 2003 4. L. Rabiner, "A tutorial on hidden Markov models and selected applications in speech recognition", Proceedings of the IEEE, vol. 77, no. 2, pp.257-287, 1989 5. K.-S. Goh, K. Miyahara, R. Radhakrishan, Z. Xiong, and A. Divakaran, Audio-visual event detection based on mining of semantic audio-visual labels, SPIE Conference on Storage and Retrieval for Multimedia Databases, Vol. 5307, pp. 292-299, January 2004 6. R. Radhakrishnan, A. Divakaran, and Z. Xiong, A time series clustering based framework for multimedia mining and summarization using audio features, ACM SIGMM International Workshop on Multimedia Information Retrieval, pp. 157-164, October 2004 7. T. Zhang and C.-C. Kuo, Audio content analysis for on-line audiovisual data segmentation, IEEE Trans. On Speech and Audio Processing, vol. 9, no. 4, pp. 441-457, May 2001 8. Divakaran, A., Miyaraha, K., Peker, K.A., Radhakrishnan, R., Xion, Z., Video mining using combinations of unsupervised and supervised learning techniques, SPIE Conference on Storage and Retrieval for Multimedia Databases, Vol. 5307, pp. 235-243, January 2004 9. T. Zhang and C.-C. Kuo. Content-based Audio Classification and Retrieval for Audiovisual Data Parsing. Kluwer Publishing Company, 2001 10. D. Reynolds and A. Rose. Robust text-independent speaker identification using Gaussian mixture speaker models, IEEE Trans. On Speech and Audio Processing, 3(1):72-82, 1995. 11. M. Casey, "Reduced-Rank Spectra and Minimum Entropy Priors for Generalized Sound Recognition", Proceedings of the Workshop on Consistent and Reliable Cues for Sound Analysis, EUROSPEECH 2001, Aalborg, Denmark, September 2001. 12. E. Scheirer and M. Slaney. Construction and evaluation of a robust multi-feature speech/music discriminator, In Proc. IEEE ICASSP, Munich, Germany, April 1997. 13. M. Rajapakse and L. Wyse. Generic audio classification using a hybrid model based on GMMs and HMMs, IEEE 11th International Conference on Multimedia Modelling, 2005. 14. C. Bishop, Neural Networks for Pattern Recognition, Oxford University Press, 2003 15. T. M. Mitchell, Machine Learning, Mc Graw-Hill, 1997 16. L. I. Smith, A Tutorial on Principal Components Analysis, Maintained by Cornell University, 2002. 17. A. Moore, Statistical Data Mining Tutorial on Gaussian Mixture Models, www.cs.cmu.edu/~awm/tutorials, CMU, 2004.